Stefan Teipel, MD1, Alexander Drzezga, MD2, Michel J. Grothe, PhD1, Henryk Barthel, MD3, Gaël Chetelat, PhD4, Norbert Schuff, PhD5, Pawel Skudlarski, PhD6, Enrica Cavedo, MS7,8, Giovanni B. Frisoni, MD7,9, Wolfgang Hoffmann, MD10, Jochen René Thyrian, PhD10, Chris Fox, MD11, Satoshi Minoshima, MD, PhD12, Osama Sabri, MD3, Andreas Fellgiebel, MD13
1Department of Psychosomatic Medicine, University of Rostock; and DZNE, German Center for Neurodegenerative Diseases, Rostock/Greifswald, Rostock, Germany
8Sorbonne Universités, Université Pierre et Marie Curie, Paris 06, Institut de la Mémoire et de la Maladie d’Alzheimer (IM2A) & Institut du Cerveau et de la Moelle épinière (ICM), UMR S 1127, Hôpital de la Pitié-Salpétrière Paris & CATI multicenter neuroimaging platform, France
9Memory Clinic and LANVIE - Laboratory of Neuroimaging of Aging, University Hospitals and University of Geneva, Geneva, Switzerland
10Institute for Community Medicine, University of Greifswald; and DZNE, German Center for Neurodegenerative Diseases, Rostock/Greifswald, Greifswald Germany
11Dementia Research Innovation Group, Norwich Medical School, Faculty of Medicine and Health Sciences, University of East Anglia, Norwich, UK
12Neuroimaging Department of Radiology, University of Utah, Salt Lake City, UT, USA
13Department of Psychiatry, University Medical Center of Mainz, Mainz, Germany
Stefan J. Teipel, M.D.
Department of Psychosomatic Medicine
University of Rostock,
and DZNE Rostock,
Gehlsheimer Str. 20,
18147 Rostock, Germany
Alzheimer’s disease (AD) is a progressive neurodegenerative disease that typically manifests clinically as a relatively isolated amnestic deficit which evolves into a characteristic dementia syndrome as the disease progresses. Advances in neuroimaging research allow mapping diverse molecular, functional, and structural aspects of AD pathology in ever increasing temporal and regional detail. There is accumulating evidence that the distinct types of AD-related imaging abnormalities follow a rather consistent trajectory during AD pathogenesis, and that first alterations may be detected years before the disease manifests clinically. These observations have fueled clinical interest in using AD-specific imaging markers to predict future development of dementia in at-risk subjects. This article summarizes current knowledge regarding the pathological validity and potential clinical utility of the best investigated imaging markers for AD. It further discusses the necessity of additional work before promising imaging markers may be successfully translated from research settings into clinical practice in routine care.
According to diagnostic criteria, such as National Institute of Neurological Communicative Disorders and Stroke and the Alzheimer Disease and Related Disorders Association (NINCDS-ADRDA) Criteria, as well as recommendations from National Institute on Aging-Alzheimer's Association (NIA-AA) workgroups on diagnostic guidelines for Alzheimer's disease, clinically early stages of AD include subjects with AD dementia as well as prodromal AD, i.e. non-demented subjects with episodic memory impairment or mild cognitive impairment (MCI) and positive imaging or neurochemical biomarkers of disease.1, 2 A more detailed outline of the concept of MCI and its subcategories can be found in the supplemental material section. While these concepts were originally designed for research purposes, they have already now begun to influence clinical practice. MCI patients who consult a specialized memory clinic are already informed about a probably underlying AD pathology and probable short term dementia development if diagnostics reveals a combined positivity of an Aβ biomarker and a biomarker of neuronal injury.2 In our view, this practice requires careful individual counselling of MCI patients at a memory clinic prior to the performance of any further diagnostic procedure. The Alzheimer’s Association goes even beyond the use of such markers in specialized care by stating: “Core clinical diagnostic criteria spelled out in the guidelines for Alzheimer's dementia and MCI due to Alzheimer's can be used now in general practice.” (http://www.alz.org/research/diagnostic_criteria/#use) Here we respond to these developments by critically summarizing current evidence for the use of imaging as a diagnostic or prognostic biomarker for both research and care settings. For the potential prognostic use of CSF markers, we refer the interested reader to recent overviews, such as provided by3, 4.
A diagnostic AD biomarker should reflect AD pathomechanisms in the presence of clinical symptoms of dementia, while a prognostic biomarker focuses primarily on the prediction of cognitive decline and dementia in prodromal stages of AD, especially amnestic MCI. Developing a valid imaging biomarker is a multi-step process which starts with methodological studies and progresses through studies involving selected samples towards studies in a real world care setting. Evidence is already available on the utility of imaging biomarkers in highly selected samples. The final proof of usefulness, however, will be the utility of imaging biomarkers in supporting a diagnosis of AD in unselected subjects from routine care, and their potential to provide cost-effective, clinically relevant outcome improvements in real world care systems. We therefore explore whether the use of novel imaging biomarkers for the detection of dementia and prodromal stages of AD, including its use to predict short to midterm conversion to dementia within one to two years, in well-defined research settings can be translated in the foreseeable future into a useful tool for improving the care of patients and their families.
Imaging markers of Alzheimer’s disease used in a research setting
The most widely used imaging modalities for the assessment of AD-related brain changes in the research setting are PET-based amyloid markers, which detect AD-specific molecular pathology, and neuronal injury markers, including structural MRI, FDG-PET, and diffusion tensor imaging (DTI), which are more robustly associated with clinical symptoms as quantified by psychometric tests compared to amyloid imaging but are only indirectly linked to the underlying molecular pathology.5 Functional MRI techniques are geared towards detecting pathological or compensatory alterations of functional network organization that may occur even before alterations in local injury markers become evident.6
Studies in selected samples have shown that AD-specific imaging abnormalities, such as cortical amyloid deposition,7 grey matter atrophy,8 hypometabolism,9 and structural10 and functional6 cortical disconnection can reliably be used to differentiate AD patients at mild to moderate dementia stages both from normal aging and from other neurodegenerative dementias. It may even be possible, using imaging assessments of the individual’s brain pathological state, to predict the risk of future development of AD dementia in a person with MCI. The suitability of imaging markers to predict conversion to AD from MCI is currently being explored in clinically highly selected samples, while the actual utility of these imaging markers in routine care will be an important topic of future research (see section 4).
The validity of an imaging biomarker can be assessed according to two principal criteria: Firstly, its pathological validity, i.e. whether or not it actually measures the pathology we expect it to measure; secondly, its clinical validity, i.e. the accuracy with which it can actually predict an individual clinical outcome. Reporting levels of diagnostic or prognostic accuracy needs to be further developed where the majority of studies does not use methods of cross-validation to assess the precision of parameters of accuracy, recommended in the statistical learning literature 11.
The following section will look at the most commonly used imaging techniques in light of current knowledge regarding their pathological validity and their value for predicting conversion to AD dementia in subjects with MCI. This encompasses advanced techniques that have some evidence of being able to reflect relevant pathogenetic events in AD and that are at least on the verge of being implemented in large multicenter diagnostic studies, such as ADNI12, AddNeuroMed13 or the European multicenter PET study14, and/or treatment trials, such as NCT01953601 (“Efficacy and Safety Trial of MK-8931 in Participants With Prodromal Alzheimer's Disease”, the APECS trial),, NCT01767311 (“A Study to Evaluate Safety, Tolerability, and Efficacy of BAN2401 in Subjects With Early Alzheimer's Disease”) or NCT01677572 (“Multiple Dose Study of BIIB037 in Participants With Prodromal or Mild Alzheimer's Disease”), available at clincialtrials.gov. We do not discuss other promising imaging technologies, such as Tau-PET (for a recent review on this technology see 15), as these techniques are undergoing dynamic development, but are not yet on the verge of major multicentre trials.
Visual analysis of amyloid-PET data provides binary information on the presence or absence of brain amyloid load (Figure 1).16 Whether or not in clinical applications quantitative approaches will have added value relative to qualitative visual analysis is still not clear. This would be the case if, for example, high amyloid levels (beyond a mere threshold of positivity), or certain regional patterns of amyloid load turned out to predict faster cognitive decline.17, 18 In multi-centre settings, the reliability of amyloid-PET binary reads across sites was found to be quite high.19 In addition, the Centiloid project aims at unifying the quantitative outcome parameters for direct comparisons between different amyloid-PET tracers.20
11C-labeled Pittsburgh compound B (PIB) and 18F-labeled amyloid-PET tracers bind with high affinity to the beta-sheets of fibrillar amyloid. They are specific in their binding to amyloid aggregates, and do not bind to other pathological proteins associated with neurodegenerative disorders, such as tau or alpha-synuclein.21
In a sample of 15 end-of-life dementia and non-dementia cases, a binary read of PIB reached a pooled sensitivity of 73% and pooled specificity of 100% in relation to the histopathological presence or absence of amyloid.22-28
PIB uptake in vivo yielded a mean (over all PIB publications available in the literature) correlation coefficient of r = 0.88±0.04 with the number of amyloid plaques upon autopsy.29-31 The 18F-labeled amyloid-PET tracers had a pooled sensitivity and specificity in 252 cases of as high as 92% and 95%, respectively,32-37in their detection of amyloid aggregates. These cases covered end-of-life dementia and non-dementia individuals in which in vivo PET imaging was compared to post mortem histopathology, as well as normal hydrocephalus patients scheduled for shunt surgery in which in vivo PET imaging was compared to ex vivo histopathology of the respective biopsy specimens. Association analyses between tracer uptake and histopathology revealed a mean (over all 18F-labeled amyloid-PET tracer publications available in the literature) correlation coefficient of r = 0.69±0.07.16, 35 One reason for this correlation being less close than in the case of PIB might be the higher unspecific white matter uptake of 18F-labeled tracers, which in turn could cause spillover effects that compromise the evaluation of grey matter amyloid load.
In most of the above studies, PET imaging was carried out in an end-of-life situation, i.e. in a scenario which might compromise the ability of PET tracers to determine brain amyloid load due to advanced brain atrophy and blood circulation impairments. Nevertheless, the amyloid-PET tracer validity data currently available are deemed satisfactory by the scientific community and the regulatory authorities.38
Potential for clinical application
According to Johnson et al. (Panel 1),39 the main role of amyloid-PET in clinical routine is in the early stages of dementia. One particularly attractive approach would be to use amyloid-PET imaging to predict, on an individual basis, the risk of conversion to AD in subjects with MCI. Current knowledge about the capacity of amyloid-PET in this regard is summarised in Table 1. While early data suggest that amyloid-PET might be useful for predicting conversion to AD at the MCI stage, it is arguable, whether the speed of possible conversion to AD in case of MCI is adequately addressable by amyloid-PET alone. As amyloid accumulation presumably represents an early event in the amyloid cascade model and precedes cognitive decline, it may have already reached a plateau in many subjects with MCI. For predicting the speed of conversion, then, neuronal injury biomarkers such as MRI, FDG-PET or CSF tau would seem to be more powerful. Studies into amyloid-PET use in primary care still need to be carried out.
FDG-PET detects patterns of altered regional glucose metabolism associated with various dementing disorders. FDG-PET results are highly reproducible between different scanners, centers and readers. As demonstrated in a recent study including 50 PET centers with 15 different scanner models, specific filtering and smoothing procedures may contribute to a further reduction of inter-scanner differences in the FDG-PET data.40 Diagnostic interpretations are commonly performed by visual inspection. The use of semi-quantitative, statistical mapping analyses has been shown to improve differential diagnostic accuracy41 and the detection of subtle metabolic changes in early AD or MCI (Figure 2).42
The cellular mechanisms of glucose consumption by neurons have been the subject of a number of investigations. One working model considers glucose metabolism in the context of capillary-glia-neuron coupling at glutamate synapses.43 Reduced synaptic metabolism of glucose (and FDG uptake) coincides with the synaptic vulnerability seen in early AD, e.g. in predicting conversion to AD in MCI patients (see Table 1).
Several retrospective investigations have employed autopsy confirmation of FDG-PET findings: FDG-PET reached a sensitivity of 84% to 96% and a specificity of 73% to 74% in the prediction of AD pathology upon autopsy in two studies involving more than 100 cases with AD, more than 25 cases with other forms of neurodegeneration and more than 20 cases without neurodegeneration (all examined clinically and neuropathologically).44, 45 On the basis of autopsy diagnosis, FDG-PET could be shown to have clearly outperformed the initial clinical diagnosis: For example, whereas the clinical diagnosis of AD was associated with a 70% probability of detecting AD pathology it increased to 84% with a positive PET-scan and decreased to 31% with a negative PET-scan.38
Potential for clinical application
Key studies into the use of FDG-PET to predict conversion to AD among MCI patients are listed in Table 1. A meta-analysis involving six studies demonstrated sensitivity and specificity of 89% and 85%, respectively, in predicting conversion. As compared to structural MRI, FDG-PET performed statistically better in positive likelihood ratio and odds ratio.46 Another meta-analysis involving seven studies demonstrated sensitivity and specificity of 77% and 74%, respectively, for FDG-PET, in comparison to 93% and 56%, respectively, for amyloid-PET.47 In a multi-modal imaging study, a direct comparison between FDG-PET, PIB-PET and structural MRI was performed in a group of MCI patients monitored longitudinally for conversion to AD. This study demonstrated FDG-PET to be more sensitive and more specific than MRI hippocampus volumetry and at least as accurate as amyloid-PET.48 These data support the notion that FDG-PET might be useful as a predictor of short term cognitive decline in clinical studies and experimental settings, but FDG-PET is still at the stage of being implemented into clinical routine, and studies in a real routine care setting are scarce (see sections 4.1 and 4.2, below).
Structural MRI is a widely established method to measure regional and global brain volumes in vivo. Volumetric MRI measurements are sensitive to variations in image contrast and resolution caused by differences in scanner settings across multiple sites. However, the multicentre variability can be effectively reduced by the use of phantom-based standardization, where a physical object with fixed geometry is scanned at each site and the imaging outcomes serve to calibrate the acquisition protocols across sites. A further step to reduce multicentre variability is the use of uniform post-acquisition image processing tools to correct for most commonly encountered image artifacts.
Brain atrophy as seen with structural MRI can, in principle, reflect the loss of tissue components, such as neurons, synapses, dendrites, as well as of glial cells and vessels. Declining hippocampal volumes in AD as determined antemortem using MRI strongly correlate with histopathological estimates of total neuron numbers upon autopsy.49 However, cell loss is not necessarily the only cause of the volumetric changes seen with MRI. The cortical thinning of the frontotemporal neocortex observed in normal aging, for instance, has been found to be primarily due to alterations in cell morphology and dendritic architecture, rather than neuronal loss.50 While medial temporal lobe atrophy on MRI is not specific for AD pathology,51, 52 the differential pattern of brain-wide atrophy can accurately separate pathologically-confirmed AD patients from healthy controls (sensitivity: 97%, specificity: 94%)53 as well as from dementias with other underlying pathologies, including frontotemporal lobar degeneration and Lewy-body disease (sensitivity: 91%, specificity: 84%)8 Potential for clinical application
Initial quantitative evaluations of structural MRI for predicting AD conversion in MCI subjects have used visual scoring scales54 or manual volume measurements of hippocampus or entorhinal cortex (Table 1), (Figure 3).55 The volumetric methods have achieved promising prediction accuracies for AD in MCI of around 80% in controlled monocentre studies. Novel computational image processing algorithms permit the automated segmentation of hippocampus volumes in MRI images with high anatomic accuracy and in a time-efficient manner.76 The application of these methods to large multicentre MRI data sets yielded somewhat lower prediction accuracies of 65% to 70% (see Table 1). Markers of hippocampal atrophy do not take into account the overall pattern of brain atrophy in AD. For this reason, more comprehensive quantitative metrics of AD-typical brain damage have been developed which use composite measures across several AD-susceptible brain regions56 or advanced machine learning algorithms to detect multivariate patterns of brain atrophy.57 These methods achieved accuracies ranging from 65% to 80% correct group separations between MCI converters and non-converters in large-scale multicentre datasets (see Table 1). A major determinant of the prediction accuracy achieved appears to be the time to conversion of the sample analysed.58 Atrophy-based classification models can detect MCI subjects on the verge of converting to AD dementia (e.g. within the year following the MRI scan) with higher accuracy than those who will convert to AD dementia within three years or more.
Although the degree of accuracy currently achieved in predicting AD dementia in subjects with MCI has not yet reached clinically acceptable levels, MRI-based sequences may provide the clinician with complementary information of a patient’s AD risk.59 While volumetric measures on MRI may not be superior over detailed clinical evaluations in predicting progression from MCI to AD dementia,60 highest prediction accuracies are being achieved when the two types of information are combined.59, 61
DTI and rs-fMRI
Histopathological studies of AD point to demyelination and axonal damage to white matter fibre tracts, which may result in functionally relevant disconnections between brain areas in addition to localized grey matter deficits.62 DTI techniques measure the directionality of water diffusion and may detect microstructural damage to white matter fibre connections through changes in specific diffusion metrics.63 Determining the most robust acquisition parameters and processing strategies for DTI is still an active area of research, and initial clinical and physical phantom data, i.e. scans obtained from a volunteer as well as a physical object with defined diffusion properties, suggest that the variability of DTI-based diffusion metrics across a range of MRI scanners is at least 50% higher than that of volumetric measures in a direct comparison between modalities.64
As well as being traceable via assessments of structural connectivity, breakdown in functional connectivity can be measured using resting-state functional MRI (rs-fMRI) techniques.6 Rs-fMRI analyses temporal correlations in spontaneous BOLD (blood oxygenation level dependent) signal fluctuations originating in spatially segregated regions of the brain. Reproducible correlation patterns of spontaneous low-frequency fluctuations in the signal between distant brain areas determine distinct sets of what are known as intrinsic connectivity networks. While these networks show robust reproducibility within-subject over time and across multiple scanners,65 methods that would permit the precise quantification of connectivity changes and thus the design of AD biomarkers are still being explored, such as in ADNI66 or DIAN.67
Clinicopathological correlation studies have provided initial evidence of the validity of DTI metrics in depicting distinct white matter alterations in neurodegenerative conditions.68 Data from experimental animal models indicate that alterations in specific diffusion metrics may, in principle, even allow inferences to be made about the specific nature of fibre damage, such as myelin degeneration or axonal damage caused by Wallerian degeneration.63
The precise nature of functional connectivity at rest is not well understood, but the intrinsic connectivity networks thus obtained appear to reflect the repertoire of co-activated functional networks that are engaged under task conditions. Furthermore, simultaneous assessment of neuronal activity using electrophysiological recordings has provided evidence that the coordinated rs-fMRI signal fluctuations do indeed represent the synchronous neuronal firing of distant brain regions rather than effects from imaging artifact.69 Potential for clinical application
Meta-analytical studies show that DTI measurements of limbic regions discriminate between healthy controls and AD or even MCI patients, with effect sizes comparable to hippocampus atrophy (Figure 4).70 DTI has recently entered the phase of large-scale multicentre studies, but its clinical utility in predicting progression from MCI to AD has only been assessed in a few monocentre studies so far (Table 1). Though promising, the high multicentre variability of this complex imaging technique may impose important constraints on its application in the wider clinical context.
The intrinsic connectivity network most intensively studied in AD is the default mode network (Figure 5),6 which consists of regions that show a striking resemblance to typical predilection sites of AD pathology. These include the posterior cingulate, medial temporal, temporoparietal and medial prefrontal cortex. Although numerous rs-fMRI studies have demonstrated that a specific disconnection between default mode network components occurs early in AD, only very limited evidence of the clinical utility of rs-fMRI in predicting AD conversion in MCI subjects is available to date, stemming from relatively small monocentre studies (Table 1).